# Bubble AI App Builder: Generation, Agent 2 and AI Credits

> How Bubble AI drafts an app from a prompt without credits, what Agent 2 can and cannot edit, how AI credits differ from workload units, and the Bubble MCP (beta).

_Source: https://professionalstoolkit.com/articles/bubble-ai-app-builder — The Professional's Toolkit · updated November 2026_

---

Bubble (bubble.io) builds a first draft of your app from a prompt at no credit cost, then charges AI credits when its Agent edits that app. Every plan carries a monthly Agent allowance per app: 500 credits on Free, 1,000 on Bubble's Starter plan, 2,000 on Growth and 4,000 on Team, the same on every tab and billing term [1]. The working rule follows. Let generation lay down the skeleton, make small fixes by hand in the visual editor, which costs nothing, and save credits for multi-step changes. Then read every privacy rule the AI wrote.


**0 credits** — What generating a new app costs in AI credits, as Bubble's pricing FAQ reads in November 2026; Agent edits draw on the app's monthly allowance.



- **500-4,000** — AI credits a month per app, Free to Team plan
- **2,500** — Characters allowed in the AI prompt box
- **6 months** — Life of purchased AI credit packs (paid plans only)


What comes out is not code. Bubble AI writes Bubble's own visual logic: pages, data types, workflows and privacy rules you can open and change in the editor. You can inspect everything the AI did without reading JavaScript, but nothing it builds leaves Bubble as source code, because apps run only on Bubble's platform. Bubble brings AI to a visual platform that, by Bubble's own account in February 2026, has been around for 14 years.

## From one prompt to a working first draft

Generation starts from Create with AI on the dashboard. You describe the app in a prompt box capped at 2,500 characters, answer multiple-choice clarifying questions, then review a plan of features and user stories beside a live preview before clicking Generate. Color codes or an inspiration image can steer the style. Bubble's release notes give build times of 10 to 15 minutes, longer for complex prompts (September 2026). The draft holds pages, a database with sample data, workflows, privacy rules per data type and complete sign-up, login and password-reset flows.

Bubble suggests the prompt cover the kind of app, its audience, its core features and your design preferences; it rejects prompts it judges malicious or too vague. Because generation spends no credits, a weak draft is cheap to discard, and Bubble's November 2026 blog says this holds for up to 10 new apps a month. **Spend your effort on the prompt.** Native mobile app generation still carries a beta label, so expect more hand work on a phone project.



> ⚠ **Free apps cannot buy packs:** AI credit packs are one-time, non-refundable, for paid plans only, and expire six months after purchase; a Free app waits for the 1st, upgrades or, if created within the last 10 days, starts a Starter trial.

## Agent 2 edits; the hard parts stay yours

Once a draft exists, the AI Agent takes over in a chat panel. Agent 2, a full rebuild that left beta in October 2026, works on existing apps too. It edits pages and elements, structures data types, writes expressions, builds frontend and backend workflows, creates API Connector calls from a plain-language description, and installs free or already-purchased plugins, favoring Bubble-built and certified ones. Bubble says Agent 2 is one and a half times better at building what was asked and twice as good at a fully functional result (October 2026).

The manual is plain about where it stops. The Agent declines very complex workflows and cannot build custom events, database trigger events, or payment and analytics actions in backend workflows; for those it explains the setup instead. As of January 2026 it wrote no custom code, and Bubble's co-CEO said in August 2026 that it still needs to catch up with tools like [Lovable](/articles/lovable-review) in conversation. **Plan to finish billing logic and triggers yourself.** Everything it builds stays visual and editable by hand.

Control comes from three approval modes: Ask confirms every change, Auto is the default for new chats, and Approve applies changes without waiting [2]. The AI tab in Settings sets the default and decides whether the Agent may read app logs and data, a right limited to owners and admins. The Agent follows the permissions of whoever opens the chat. Several chats can run as tabs, voice prompts work, and it checks its changes with the issue checker and screenshots.


> 💡 **Set Ask mode on live apps:** In the AI tab of Settings, make Ask the default approval mode for apps with real users, so every Agent change waits for your confirmation.


## Stretching each month's credit allowance

Credits belong to the app, not your account. Everyone editing one app draws on a single balance that cannot be pooled or moved, and included credits refresh on the 1st without rolling over. Bubble sets no fixed cost per request: usage is worked out after it runs, grows with size and complexity, and per-action amounts can change [3]. So any budget is an estimate. The credit breakdown view, added in October 2026, lists each lot with its balance and expiry date.

| Plan | AI credits a month, per app | Workload units a month, per app |
|---|---|---|
| Free | 500 | 50K |
| Starter | 1,000 | 175K |
| Growth | 2,000 | 250K |
| Team | 4,000 | 500K |

Running dry mid-task is not a disaster. The Agent stops, keeps what it finished, says what did not run and charges nothing automatically; the visual editor keeps working. Paid apps can add one-time packs: 1,000 credits cost $10 and 10,000 credits cost $100, which by our arithmetic is $0.01 per credit at every size, so a bigger pack saves nothing. Purchased credits expire six months after purchase. A Free app cannot buy packs; it waits for the 1st, upgrades, or starts a Starter trial.

That Starter trial lasts 14 days, must start within 10 days of creating a new app, takes a $1 card pre-authorization and converts to paid Starter unless canceled; neither a converted trial nor a credit pack is refundable. Cheaper still is the credit you never spend. Moving a button or renaming a field costs nothing in the visual editor, while an Agent turn bills for the same change, and image, page and test data generation draw on credits too. Keep Agent turns for multi-step work, such as a new page with its workflows and data, where it saves the most clicking.

## Credits build the app; workload units run it

Bubble meters two different things. AI credits pay for building with the Agent, page and image generation and test data. Workload units (WU) pay for the server work of running the app: searches, writes, API calls, scheduled workflows, file uploads and emails. They are separate allowances, running out of one does not touch the other, and neither converts into the other. The Development version counts toward WU as well, so test sessions are metered too.

That matters because a generated app has real server logic. Backend workflows the Agent writes, recurring jobs it schedules and API calls it wires up all spend WU once they run. App metrics in the Logs tab shows which pages, workflows and expressions used the most after a test session. Past the plan's monthly WU, paid apps pay overages at $0.30 per 1K WU, on by default, or buy a workload tier; turning overages off takes the app offline at the limit until overages are re-enabled, a tier is bought or the next billing period starts, as the [pricing breakdown](/articles/bubble-pricing) explains.


> ⚠ **Credit use per request varies:** Bubble sets no fixed credit cost per Agent request and can change per-action amounts, so treat any monthly credit budget as an estimate.


## Reading what the AI built before launch

Generated privacy rules are a starting point, and Bubble asks owners to review them before going live. They are server-side conditions on each data type and field; without them any data sent to a device can be read, and hiding an element protects nothing. Check who may find, view and search each type the Agent created. Since October 2026 the preview tab can run the app as any test user. Our [database and workflows guide](/articles/bubble-database-and-workflows) walks through the rules field by field.

Next, run the security dashboard: five basic tests on Bubble's Starter plan, advanced checks and unlimited tests from Growth, automation and scheduling on Team. It flags problems but fixes nothing, and Bubble's SOC 2 claim does not make your app compliant. Bubble's preliminary root-cause analysis (November 2026) of the October 2026 incident says the attacker got partial access to Bubble's production systems and some user data, including private plugin data such as credentials plugin developers had saved, and targeted individual apps through platform issues that reached stored credentials or data even in correctly set-up apps, and through settings such as privacy rules that granted access; affected owners were contacted. Bubble advises reviewing privacy rules and marking API keys private. A final RCA will follow.

Going live is its own gate. A Free app has no live version, so publishing needs a paid plan: Bubble's Starter plan (Web) is $32 per month billed annually or $39 per month billed monthly, per app. A native app needs Bubble's Starter plan on the Mobile tab ($42 per month billed annually or $49 billed monthly, per app) or the Web & Mobile tab ($59 per month billed annually or $69 billed monthly). Deploying moves structure, not data, so sample records stay in Development unless you copy them. A web deploy also needs the issue checker at zero issues. Clear them early.


> 💡 **Preview as a stranger first:** Run the app in the preview tab as different test users and confirm the generated privacy rules hide the fields they should.


## Bringing an outside agent in through the Bubble MCP

Builders who work in Claude, Claude Code, ChatGPT, Codex or GitHub Copilot in VS Code can drive a Bubble app from there through the Bubble MCP, in beta since October 2026. The connected agent builds with ordinary Bubble elements, workflows and data types, so the result stays visually editable. It can read live app data and create backend workflows and API calls, for web and native mobile. At launch it could not merge branches or deploy. Access goes through OAuth and can be revoked.

The cost model is the draw. MCP work uses no Bubble AI credits and is available on every plan, free during the beta; the AI work counts against your plan with the AI provider, whose limits still apply [4]. You also pick the model, whereas Bubble AI uses Bubble's own mix of models per task, with Claude Sonnet 5 across the editor since September 2026. One point cuts the other way. Bubble's Terms let it train its AI tools on their input and output, and Bubble publishes no opt-out.

Some posts describe the MCP as a way out. Read that carefully. An outside agent can read an app's data types, pages, workflows and privacy rules and recreate it in code, without guaranteed fidelity, while API keys must be copied by hand and hosting, database and sign-in become your job. That is a rebuild, not a code export, which Bubble does not offer.


## FAQ

**Can my finished Bubble app call OpenAI, Anthropic or Gemini models?**

Yes. Any AI provider with an API connects through plugins or the API Connector, including streaming responses, and the provider bills you separately. Bubble had no native LLM action as of October 2026; it is planned.

**Who can buy extra AI credits for a Bubble app?**

Any collaborator with editor access can use the app's credits, but only the app owner or an admin can buy more, and only on a paid plan.

**What happens to purchased AI credits if I downgrade to Free or cancel?**

They stay in the app until their six-month expiry. Included credits switch to the new plan's allowance without proration, and a suspension for non-payment forfeits unused credits.

**Can I bring an app built with another AI builder into Bubble?**

Universal import, in beta for a subset of users, takes a codebase ZIP and rebuilds it as a visual Bubble app; Bubble lists exports from Lovable, v0, Base44, Replit and Claude as working well. Exports are often incomplete, so API keys, data or outside services may need to be set up again.


## The bottom line

Bubble AI works best as a drafting tool on top of a visual platform you will keep editing by hand. Generation costs no credits, so iterate on the prompt until the draft fits. Spend the allowance, 500 to 4,000 credits per app a month depending on plan, on multi-step Agent work, and keep cosmetic fixes in the editor. Budget workload units separately. Pick it if you want AI help without a codebase to maintain; skip it if you need exportable code. Either way, review the generated privacy rules before launch.



## Sources

[1] Bubble Pricing — Start Free, Scale as You Grow — https://bubble.io/pricing (2026-11)
[2] Bubble Manual: Edit apps with AI — https://manual.bubble.io/help-guides/ai/bubble-ai-agent/edit-apps-with-ai (2026-11)
[3] Bubble Manual: AI credits — https://manual.bubble.io/help-guides/ai/ai-credits (2026-11)
[4] Bubble Manual: Bubble MCP — https://manual.bubble.io/help-guides/ai/bubble-mcp (2026-11)
[5] Bubble Terms of Service — https://bubble.io/terms
[6] Bubble Privacy Policy — https://bubble.io/privacy